This is the PyTorch implementation of our SIGIR 2022 paper RankFlow: Joint Optimization of Multi-Stage Cascade Ranking Systems as Flows.
The raw datasets are ML-1M, TianGong-ST, Tmall.
The downloaded raw data should be placed into [dataset_name]/raw_data folder. And all the preprocessed data will be placed at [dataset_name]/feateng_data folder. To execute the data processing procedure, just use
python process_[dataset_name].py
in the corresponding directory.
After the preprocessing, execute the dataset generation process in code folder as
python dataset.py -d [dataset_name]
To execute the independent training on the impression (displayed) data, use
python warmup.py -d [dataset_name] -m [model_name]
or use the shell scripts (to train multiple models) as
bash ind_train.sh -d [dataset_name] -m [list of model_names]
To execute the RankFlow joint training, use
python train.py -d [dataset_name] -ms [list of model_names in the cascade]
@inproceedings{qin2022rankflow,
title={RankFlow: Joint Optimization of Multi-Stage Cascade Ranking Systems as Flows},
author={Qin, Jiarui and Zhu, Jiachen and Chen, Bo and Liu, Zhirong and Liu, Weiwen and Tang, Ruiming and Zhang, Rui and Yu, Yong and Zhang, Weinan},
booktitle={Proceedings of the 45th International ACM SIGIR Conference on Research and Development in Information Retrieval},
pages={814--824},
year={2022}
}